When a CRISPR screen uses standard Cas9, every guide cut is also a wound: the resulting double-strand break triggers the cell's p53-guarded DNA damage response, which arrests or kills healthy cells while quietly sparing — and thereby enriching — cells whose p53 pathway is broken. Nathan H. Lazar et al. showed just how far this distortion reaches: Cas9 cuts in primary human endothelial cells frequently delete everything from the cut site out to the chromosome end, so two unrelated genes on the same chromosome arm end up with nearly identical measured phenotypes (within-arm cosine-similarity ranking probability 0.71, versus a 0.50 null) (High-resolution genome-wide mapping of chromosome-arm-scale truncations induced by CRISPR–Cas9 editing). This matters beyond screen hygiene: the RxRx3 phenomics resource, which matches CRISPR knockout morphologies against 1,674 small-molecule perturbations to predict drug-target interactions, is built entirely on those same confounded Cas9 knockouts (RxRx3: Phenomics Map of Biology).
The proposal here is to rebuild the genetic side of that drug-target matching pipeline on single-nucleotide base editors — engineered Cas scaffolds fused to deaminase enzymes that chemically convert individual DNA letters (for example C-to-T) without ever cutting both strands (CRISPR technologies for genome, epigenome and transcriptome editing) — and to pair it with single-cell measurements of both RNA and protein, because transcript levels track protein abundance only 40–60% of the time (A high-stringency blueprint of the human proteome). Reviews assert that base editing "circumvents p53 activation," but no study in the literature has actually run a genome-scale phenotypic screen with base editors to verify this, nor asked how much of the current ~70% failure rate in drug-target matching is truncation artifact versus post-transcriptional noise. This project measures both.
Core Claim
Replacing Cas9 double-strand-break knockouts with cytosine base editor (CBE) nonsense knockouts — installing a premature stop codon by a single C-to-T edit — will collapse the chromosome-arm proximity bias in phenomic screens from 0.71 toward the 0.50 null baseline, and this cleaner genetic reference panel will improve zero-shot drug-target interaction matching above the current best of 0.307 average precision. Furthermore, by reading out a paired RNA-plus-protein subset, we can decompose the remaining matching error into how much comes from transcript-protein discordance versus residual editing artifacts, and use pharmacological inhibitors as an orthogonal, time-resolved comparator to validate that base-edited loss-of-function and drug inhibition of the same node land on the same phenotype.
The mechanism is concrete: a stop codon placed early in a coding sequence removes the target protein via nonsense-mediated decay and translation loss — a permanent, heritable, node-specific perturbation — without the telomeric truncations whose survival is gated by p53 status (High-resolution genome-wide mapping of chromosome-arm-scale truncations induced by CRISPR–Cas9 editing). Precedent for exactly this edit logic exists: CBE4max was used to introduce a premature stop codon in PKP2 in pigs with no collateral rearrangements (<i>PKP2</i>杂合敲除诱导的致心律失常性心肌病猪模型与表型分析).
Research Program
The program has three linked components. First, we construct an arrayed base-edited knockout library covering the 736 unblinded gene targets of the RxRx3-core benchmark in HUVEC cells — the same cell type, Cell Painting protocol, and 1,674-compound, 8-dose chemical arm used there (RxRx3-core: Benchmarking drug-target interactions in High-Content Microscopy) — so the comparison is like-for-like. Guide design uses PAM-flexible scaffolds (SpG, SpRY) to reach loci excluded by the canonical NGG rule (PAM-flexible genome editing with an engineered chimeric Cas9), and we filter guides whose deamination window contains bystander bases, using the attention-based outcome predictor of Mollaysa et al., which reports Spearman correlations of 0.85–0.89 across editor-scaffold combinations on libraries of roughly 11,000 reference sequences (Attention-based Multi-task Learning for Base Editor Outcome Prediction). Each edited clone is genotyped, so knockout status is confirmed at the DNA level rather than inferred, and we run matched arms in a p53-wild-type line and a p53-deficient line to test whether the nickase half-cut of the base editor still produces low-frequency arm loss when the damage filter is off — an open question the reviews explicitly do not answer (Harnessing the evolving CRISPR/Cas9 for precision oncology).
Second, we quantify proximity bias directly: cosine similarity between phenomic vectors of genes on the same chromosome arm versus different arms, with the Cas9 value of 0.71 and the 0.50 random expectation as the two anchors (High-resolution genome-wide mapping of chromosome-arm-scale truncations induced by CRISPR–Cas9 editing). We also assay structural variants directly, since Cas9 produced physical copy-number loss toward the telomere at 4–25% of targeted genes in that study.
Third, we retrain a Phenom-2-style self-supervised encoder (the 1.86-billion-parameter vision transformer from the RxRx3-core benchmark) on the base-edited images, with identical compute and tuning budget to its Cas9-trained counterpart, and evaluate zero-shot drug-target matching by cosine similarity against the same compound set. On a CITE-seq subset — at least 300 cells per perturbation, the threshold recommended by the scPerturb harmonization study (scPerturb: Harmonized Single-Cell Perturbation Data) — we measure RNA and surface protein jointly, so that for each gene we can compute how much the transcript-level knockout signal understates the protein-level effect. The cautionary number here is the 76.5% protein drop that coexisted with only a ~15% mRNA drop in Cas13 perturbations (Efficient combinatorial targeting of RNA transcripts in single cells with Cas13 RNA Perturb-seq); our paired readout is what lets us separate real biological discordance from measurement blind spots.
A final decisive check connects base editing to pharmacological pathway inhibition at a single, well-characterized locus: Livingstone et al. found that the NT5C3A gene switches transcription start sites under interferon treatment, creating a 5' TOP-like motif that makes its translation selectively sensitive to mTOR inhibition, one of only 3 selectively sensitive interferon-stimulated genes among hundreds, under a Torin1 regime that otherwise represses global translation by roughly 50% (Assessment of mTOR-Dependent Translational Regulation of Interferon Stimulated Genes). We will use ABE editing to install and to remove that motif, then test whether motif-edited cells gain or lose Torin1-selective (versus rapamycin) translational sensitivity. This is an experiment no small molecule and no Cas9 knockout can do: it perturbs a cis-regulatory nucleotide, not a protein, and no blunt deletion could isolate a six-base motif.
Baselines
The naive reference point is the uncorrected Cas9 status quo: Cell Painting features from the standard Cas9 knockout library, which the RxRx3-core benchmark scores at 0.276 average precision for drug-target interaction matching with CellProfiler-derived features (RxRx3-core: Benchmarking drug-target interactions in High-Content Microscopy). It is the floor the whole field currently stands on, and it inherits the full truncation confound.
The strongest competitor is the Phenom-2 encoder, a 1.86-billion-parameter self-supervised vision transformer trained on Cas9 knockout images, which achieves 0.307 average precision on the same benchmark (RxRx3-core: Benchmarking drug-target interactions in High-Content Microscopy). To make the comparison fair, we retrain Phenom-2 from the same architecture and data budget on our base-edited images, match the compound set, doses, and evaluation protocol exactly, and give both pipelines the same hyperparameter search. Beating it means average precision above 0.307 on the identical split, with no additional inputs beyond the cleaner edited library.
The key ablation isolates the ingredient this proposal claims matters — cleavage-free editing versus computational cleanup of cleavage-based screens. We apply the geometric subtraction correction of Lazar et al. (subtracting the average phenomic signal of unexpressed genes on the same chromosome arm) to the Cas9-trained embeddings, and compare its proximity-bias reduction and drug-target matching against the base-edited panel. If post-hoc subtraction closes most of the gap, the expensive wet-lab switch is not justified, and we would say so; if the edited panel wins, it demonstrates the bias is a physical data problem, not a modeling one. A second ablation, bystander-filtered versus unfiltered guides, tests whether editing purity or editing chemistry drives any improvement.
Why not just…?
Why not just apply post-hoc corrections to existing Cas9 screens? This is the most serious alternative, and the geometric subtraction of Lazar et al. demonstrably reduces proximity bias and improves protein-complex recall (High-resolution genome-wide mapping of chromosome-arm-scale truncations induced by CRISPR–Cas9 editing). But subtraction can only decorrelate surviving cells' features; it cannot recover the cells that p53-mediated apoptosis physically removed, nor undo the selective enrichment of damage-tolerant clones documented across therapeutic CRISPR contexts (The Application of CRISPR/Cas9 Technology for Cancer Immunotherapy: Current Status and Problems; Application and Prospect of CRISPR/Cas9 Technology in Reversing Drug Resistance of Non-Small Cell Lung Cancer). The head-to-head ablation above is designed precisely to settle this: if subtraction matches base editing on drug-target precision, the correction wins on cost, and the field should use it.
Why not just use RNA-targeting Cas13 to avoid DNA damage entirely? CaRPool-seq achieves about 80% cell assignment in triple-target screens (versus under 50% for Cas9) and never touches DNA (Efficient combinatorial targeting of RNA transcripts in single cells with Cas13 RNA Perturb-seq). But its own data expose the problem: cleaved transcripts linger as non-functional fragments, so mRNA drops ~15% while protein drops 76.5% — the readout lies about exactly the thing we want to measure. Worse, RfxCas13d collateral cleavage can cause lethal, non-specific toxicity in mammalian systems (The collateral activity of RfxCas13d can induce lethality in a RfxCas13d knock-in mouse model). Cas13 is transient and RNA-limited; base editing gives permanent, genotyped knockouts and a stable morphological phenotype for compound matching.
Why not just impute protein from RNA with existing models? Tools like totalVI (Joint probabilistic modeling of paired transcriptome and proteome measurements in single cells) and scChord (scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation) are useful, and we will use them as analysis tools. But scChord's own authors show that regression-style models over-smooth: they fit conditional means and miss bimodal distributions like CD4 and heavy-tailed markers like CD45RA. More fundamentally, imputation cannot detect discordance that is itself the signal — a knockout whose transcript is unchanged but whose protein collapses would be erased, not revealed, by an imputer trained to make RNA predict protein.
Why not just use the small molecules and skip genetics? Inhibitors are fast, reversible, and dose-tunable, and they resolved clean epistatic orders in classic pathway work (for example Torin1 versus rapamycin separating mTORC1 catalytic output from allosteric control). But they carry off-target polypharmacology, cannot touch non-druggable scaffolding proteins, and cannot edit a cis-regulatory motif — no inhibitor exists that specifically toggles the NT5C3A 5' TOP-like start-site switch. The two are complementary here: drugs give the time-resolved perturbation signature, base editing gives the permanent node- and nucleotide-level ground truth, and the final experiment tests whether they agree.
Proposed Experiments and Decisive Checks
Three outcomes decide the proposal. First, proximity bias under base editing should fall from the Cas9 value of 0.71 toward 0.50; any residual value significantly above 0.50, especially in p53-deficient cells, would implicate the base editor's nickase activity as a low-level break source and itself be a publishable safety finding — no reported result found on Lacuna for genome-scale proximity bias under base editing. Second, drug-target interaction average precision on the RxRx3-core protocol should exceed the Phenom-2 Cas9 baseline of 0.307 (with CellProfiler's 0.276 as the floor), using identical compounds, doses, splits, and compute. Third, in the CITE-seq subset, we report for each gene the ratio of protein-level to RNA-level effect size and test whether genes with high discordance are disproportionately the ones where the Cas9-trained embeddings failed to match their inhibitors — which would tie the ~70% unexplained variance in current matching to a measurable, mechanistic cause. The NT5C3A motif experiment closes the loop: ABE-mediated motif installation and removal should respectively create and abolish Torin1-selective translational repression relative to rapamycin, reproducing pharmacologically the regime Livingstone et al. discovered by inhibition alone (Assessment of mTOR-Dependent Translational Regulation of Interferon Stimulated Genes). If all three hold, the field gets a reusable, cleavage-free genetic reference panel and a quantified account of why its Cas9 predecessor was biased.
References
- High-resolution genome-wide mapping of chromosome-arm-scale truncations induced by CRISPR–Cas9 editing
- RxRx3: Phenomics Map of Biology
- Author Correction: CRISPR technologies for genome, epigenome and transcriptome editing
- A high-stringency blueprint of the human proteome
- <i>PKP2</i>杂合敲除诱导的致心律失常性心肌病猪模型与表型分析
- RxRx3-core: Benchmarking drug-target interactions in high-content microscopy
- PAM-flexible genome editing with an engineered chimeric Cas9
- Attention-based Multi-task Learning for Base Editor Outcome Prediction
- Harnessing the evolving CRISPR/Cas9 for precision oncology
- scPerturb: Harmonized Single-Cell Perturbation Data
- Efficient combinatorial targeting of RNA transcripts in single cells with Cas13 RNA Perturb-seq
- Assessment of mTOR-Dependent Translational Regulation of Interferon Stimulated Genes
- The Application of CRISPR/Cas9 Technology for Cancer Immunotherapy: Current Status and Problems
- Application and Prospect of CRISPR/Cas9 Technology in Reversing Drug Resistance of Non-Small Cell Lung Cancer
- The collateral activity of RfxCas13d can induce lethality in a RfxCas13d knock-in mouse model
- Joint probabilistic modeling of paired transcriptome and proteome measurements in single cells
- scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation